Sequential Adoption of Package Technologies The Dynamics of Stacked Trait Corn Adoption Ursula Aldana*, Jeremy D. Foltz*, Bradford L. Barham*, and Pilar Useche** *University of Wisconsin – Madison, **University of Florida Contact Author: Brad Barham
[email protected] May 2010 Abstract: GM corn seed companies have innovated continuously with the introduction of new traits and, more recently, with the creation of stacked varieties, which combine more than one trait. This work develops a Bayesian model of adoption dynamics that demonstrates how uncertainty with a package technology with known risk can lead to a sequential adoption pattern in which farmers adopt a single component first. We then develop a semiparametric panel data model of adoption dynamics to measure the effects of experience with single trait (non-stacked) varieties on the adoption of stacked varieties. The results underscore the importance of early experience with the non-stacked technology in the subsequent adoption of stacked varieties, i.e., a sequential adoption process. There is also evidence that farmers with more human capital tend to learn faster from own experience and that as the GM corn-technology diffusion process deepens, the importance of early experience decreases. Selected Paper prepared for presentation at the Agricultural & Applied Economics Association’s 2010 AAEA, CAES & WAEA Joint Annual Meeting, Denver, Colorado, July 25-27, 2010. Copyright 2010 by Ursula Aldana, Jeremy Foltz, Bradford Barham, and Pilar Useche. All rights reserved. Readers may make verbatim copies of this document for non-commercial purposes by any means, provided this copyright notice appears on all such copies.
Acknowledgements: This work was funded through the Program on Agricultural Technology Studies core budget, a UW-Hatch grant and the Illinois Missouri Biotechnology Alliance grant from the USDA. This work has benefited from discussions with Gustavo de los Campos, Michael Carter, Jean-Paul Chavas, Paul Mitchell, Terry Hurley, Laura Schechter, and participants at the Applied Economics Workshop at University of Wisconsin-Madison as well as comments from 2 reviewers and an AJAE editor.
Sequential Adoption of Package Technologies The Dynamics of Stacked Trait Corn Adoption Most recent models of technology adoption take the new technology to be a single uniform technology that firms or farmers decide whether to use. Many technologies in fact consist of sub-components that are available individually or jointly as a package. Genetically modified (GM) corn to date presents such a technology option to farmers, with single trait or multiple (stacked) trait varieties. This divisibility of the technology package can create a particular adoption dynamic, since the experience gained through the adoption of subcomponents will provide information on the characteristics of the package thereby encouraging or discouraging its subsequent adoption. A now old literature on green revolution technology adoption from the 1970’s and 1980’s analyzed package technologies, and identified sequential adoption patterns in which farmers adopted parts of the package before adopting the whole package (see e.g., Byerlee and Hesse de Polanco; Leathers and Smale). These works were primarily empirical, and argued that fixed costs, credit constraints, risk, uncertainty, and learning all contributed to a sequential adoption pattern. In the US GM corn market, we posit that learning and uncertainty are likely to be the most salient of these explanations. GM seeds are inherently divisible, US farmers are not likely to face credit constraints that bind the purchase of a sufficient quantity of seeds to test GM varieties, and crop insurance is available to help manage basic risks. 1
1
During the period of our data, 2000-2006, US grain farmers had historically easy access to credit at low interest rates. In addition, most farmers in our sample purchased crop insurance. Thus, while credit constraints and risk aversion may play a role in the sequencing of adoption, they are likely to be a small contributor to the issue. 1
Technology adoption has long been viewed as a process that hinges on farmer learning about uncertain and risky options (Griliches; Rogers; Feder and Slade; Feder et al.; Foster and Rosenzweig; Conley and Udry). Often the learning process about a new technology can drive the timing of adoption. As explained in Feder and Omara, uncertainty diminishes over time because of the common experience gained through the adoption of farmers in the economy and because of the expected improvement of extension services and advancement of research studies. Similarly, uncertainty decreases with direct experience with the technology or with a component of the technological package. A strand of the literature has modeled this decrease in uncertainty using a Bayesian approach (Lindner; Feder and Slade; Leathers and Smale), according to which producers update their beliefs about the distribution of the profitability of the technology, using observations on the profit achieved by other producers or themselves. From this literature, the Bayesian conceptual model by Leathers and Smale is the one that comes closest to our work in providing a reason for a sequential adoption process. Our work expands the Bayesian adoption literature by considering correlated learning across components of package technologies and in so doing provides a new logic for a sequential pattern of adoption. Previous studies of farmer adoption of GM crops focus almost exclusively on the binary decision to use GM varieties (Alexander, Fernadez Cornejo et.al., Hubbell et.al. ) or else on a one-time decision across multiple independent or correlated choices (Useche et al.). All of these GM adoption studies are cross-sectional analyses that preceded or coincided the introduction of stacked GM traits, and do not account for timing or previous experience in a systematic fashion. Thus, none of these studies explicitly account for the
2
potential sequencing of farmer adoption choices, especially the potential path of trying one or more single variety traits before adopting stacked varieties, which combine multiple traits. Examining the adoption of stacked varieties through a dynamic lens with longitudinal data can improve our understanding of the sequencing of farmer choices in this and other types of evolving package technologies. In this article, we analyze the sequential nature of the adoption of stacked GM corn varieties, which are a package of individual GM traits. The main empirical questions that we address are whether and why farmers might move sequentially from a single trait to a stacked variety. In the GM corn market during our study period there were three traits (herbicide tolerance (HT), bacillus thuringeiensis (BT) to kill corn borers, and to kill corn rootworm insects) in the market that were sold both independently and as a packages. We develop a novel theoretical framework based on the concept of Bayesian updating of the beliefs regarding the expected profits of the stacked varieties. We show that the possibility of learning about the profitability of the stacked variety through the use of a single trait variety reduces the uncertainty of stacked varieties, which tends to favor a sequential pattern of adoption. The empirical analysis exploits a survey of corn farmers from Minnesota and Wisconsin conducted in 2006. The 2006 survey included retrospective questions on the use of GM corn varieties as well as questions on the socio-economic characteristics of the farm operator and characteristics of the farm itself. The econometric estimations using these data examine the adoption of stacked varieties using semi-parametric fixed-effect regressions to explore the role of experience, as well as its interaction with education and with the stage of the technology diffusion process. The econometric method that we use
3
does not impose a parametric functional form on the relationship between the likelihood of adopting a stacked variety and the experience of the producer with a single trait GM technology. We use this method because the Bayesian model shows a non-linear impact of experience on the latent variable, the returns to adoption. This method allows us to examine how the adoption of stacked varieties evolves as experience changes and is also useful when comparing the impact of experience for different groups. The empirical results underscore the importance of early experience with the single trait technology in the subsequent adoption of stacked varieties, i.e., a sequential adoption process. They also provide evidence that the cost of learning decreases for farmers with time as the technology diffuses (Feder and Omara). Thus, increases in common knowledge provide partial substitutes for own experience in an adoption process and reduce the incentives for sequential adoption. The results also suggest that more educated farmers tend to learn faster and require fewer trials before adopting stacked varieties. The next section develops a Bayesian model of adoption dynamics for technologies with correlated outcomes that demonstrates how uncertainty with a package technology with known risk can lead to a sequential adoption pattern in which farmers adopt components first. The third section develops the semi-parametric econometric modeling approach, describes the data, and offers variable means on sequential adoption patterns for GM corn in Minnesota and Wisconsin. The fourth section presents the econometric results on how experience with single trait varieties, human capital, and common knowledge diffusion shape the adoption of stacked varieties. The final section summarizes our findings and reflects on research and market performance issues.
4
A Bayesian Model of Adoption Dynamics In this section, we develop a novel multivariate Bayesian model that explains the logic for a sequential adoption pattern for a package technology with individual components that could provide information on the value of the package. While this builds on a standard Bayesian learning framework (e.g., Anderson, Stoneman), it is to our knowledge the first model using a multivariate Bayesian framework to understand how farmers learn about the profitability of package technologies. Our model features a seed market where the farmer chooses between conventional, single GM trait, and a stack of multiple varieties. In this Bayesian model, farmers learn about the yields of the stacked variety by using a single trait variety. Farmers make choices for the current period based on their current information set, which includes past experience with a GM trait. This allows us to analyze how experience with a single GM trait changes the incentives to adopt the package given the information set available to the farmer.2 Farmer’s maximization problem We assume that farmers face a choice set that includes J types of corn seeds (e.g., Conventional (c), HT only (1), BT only (2), and stacked HT/BT (s)). To simplify the modeling, we focus on the case in which the farmer adopts trait 1, and how that might create learning about trait 2 and stacked traits. The farmer will choose the variety j*, which gives her the highest expected utility:
2
We do not explicitly model the potential forward-looking dynamics of farmer choices, the case where farmers might adopt a particular variety today because of the future value of information it might provide. This would unnecessarily complicate the modeling exercise and merely reinforce the sequential adoption logic we are seeking to demonstrate. 5
j* = arg max {EU j }
(1)
j
where EU j is the expected utility achieved when adopting variety j3. Following Stoneman, we express the farmer’s expected utility as a linear function of the expected value as well as the variance of profits4: EU j = E (π j ) − bV (π j )
(2)
where b represents the farmers risk aversion level. As detailed below, we assume that the farmer does not know the expected profits of GM varieties, and this implies that the variance and the expected value in (2) represent both the parameters of the real distribution of profits and the parameters of the farmer’s beliefs. The term E (π j )
represents the mean of the distribution of the farmer’s beliefs regarding expected profits, which may or may not be accurate to the true expected profits. The aggregate variance, V (π j ) , is the sum of two elements: the variance of the distribution of the beliefs regarding expected profits, and the variance of these profits implied by variation in common known but stochastic factors such as weather conditions. The variance of the beliefs can be interpreted as the uncertainty component while the variance of the profits implied by weather conditions represents what is typically thought of as production risk.
3
In this model we constrain farmers to choose one variety at a time to learn from. An alternative specification would assume low sunk/fixed costs to adoption and have the farmer choose a land portfolio in order to maximize her expected utility. Specification (1) simplifies the model to focus the analysis of the impact of experience. The same results as those presented here can be obtained from a portfolio selection model, but they would also be a function of the amount of land chosen for each crop type. Since the latter would be endogenously determined in the model, it would add additional conditions and complicate signing the direction of learning effects.
4
This standard mean-variance functional form is consistent with a farmer who has a CARA utility function when the actual profits and the farmer beliefs regarding the mean profits are normally distributed. 6
Profits depend on the output price, the yield, the input cost as well as the price of a particular seed variety in the following manner:
π j = p * y j − w j − p sj , m
(3)
where p m is the output price, y j is the yield achieved by variety j, w j represents the input costs incurred when variety j is used and p sj is the price of seed of variety j. We assume the yields, as well as the input costs are stochastic and depend on the variety chosen. Input costs are distributed normally, with a mean value of μ w j and a variance of
σ w2
j
, both of which we assume are known by the farmers. The distribution of yields is
described below. The yields achieved for each type of variety depend on the traits included in a particular seed. Define Ijk as an indicator function that equals one if variety j contains trait k, and define Ic as an indicator function that equals one if variety j is the conventional variety. Assuming that the number of traits is equal to 2, the yields can be expressed as: y j = δ c I c + δ1 I j1 + δ 2 I j 2 + δ12 I j1 I j 2 ,
(4)
Equation (4) indicates that the yield of a conventional variety is equal to δ c , while the yields of GM varieties depend on the traits contained in the seeds. In this formulation
δ12 captures the impact on yields of having more than one trait present in a seed variety. We assume that δ c δ1 and δ 2 are random variables which are distributed as N ( μi , σ i2 ) ,
where i=c,1,2; while δ12 is a random variable distributed as N ( μ12 , σ 122 ) ; and that these distributions are independent from each other. Note that the parameters σ c2,σ12 ,σ 22 ,σ122 represent production risk.
7
Farmer Learning about the Technology
In order to capture the uncertainty in the profitability of a new technology, we assume that the farmer knows the expected value of the yield of the conventional variety, μc , the variance of all yields ( σ c2,σ12 ,σ 22 ,σ122 ), but does not know the expected values of the yields of the GM varieties: μ1 , μ2 and μ12 . Instead, the farmers have a prior distribution, which captures their expectations regarding the value of these parameters. This prior distribution is updated in a Bayesian manner into a posterior distribution, as the farmer experiments with the technology. We conceptualize the package as having two correlated elements [ μ1 , μ2 ] that can be learned about through experience with any of these elements and then a third part, μ12 , which is the part of the package that can only be learned about through use of the package.5 In this conceptualization, we assume that the prior distribution of [ μ1 , μ2 ] is bivariate normal with a positive covariance between the two varieties. This captures the idea that experience with one trait provides information on the qualities of the other trait. For example, if the same company produces both traits, the farmer is likely to expect better (worse) results for trait 2 when the results for trait 1 turn out well (badly). Note that it is possible for the prior distribution of [ μ1 , μ2 ] to have a positive covariance while δ1 and δ 2 are independent from each other. The assumption of a positive covariance between μ1 and μ2 means that the farmer believes that if the expected value
Our assumption that μ12 is uncorrelated with the other elements of the package is a simplifying assumption. The sequential learning results would only be reinforced if we allowed μ12 to be correlated with μ1 and μ2 (i.e., learning about all elements of the package).
5
8
of trait 1 yields is high, then the expected value of trait 2 yields is more likely to be high. The assumption of independence between δ1 and δ 2 establishes that, regardless of the expected value of the distribution of [ δ1 , δ 2 ], a good year for trait one does not imply that a good year for trait 2 is more likely to happen. On the other hand, the prior distribution of
μ12 is assumed to have a mean given by θ120 and a variance given by V120 . In order to simplify the analysis, we assume that beliefs regarding μ12 are uncorrelated with beliefs regarding [ μ1 , μ2 ]. With the Bayesian formulation, the prior distribution of [ μ1 , μ2 ] is denoted as p o ( μ1 , μ2 ) and is described as:
⎡⎛ μ ⎞ ⎛ θ 0 ⎞ ⎤ p o ( μ1 , μ2 ) = N ⎢⎜ 1 ⎟ ⎜ 10 ⎟ , Σ ⎥ , (5) ⎢⎣⎝ μ2 ⎠ ⎝ θ 2 ⎠ ⎥⎦ where [ θ10 , θ 20 ] are the expected value of the distribution and ∑ is the variance covariance matrix which equals: ⎛V0 ∑ = ⎜ 10 ⎝ C12
C120 ⎞ ⎟ V20 ⎠
.
(6)
In order to illustrate the sequencing of adoption we assume that the farmer starts by growing a single trait variety that contains trait 1. She uses n observations (e.g., years growing the crop) on the yields of a single trait variety containing trait 1 to update her beliefs regarding μ1 . Given the assumption of a positive covariance between μ1 and μ2 in the prior distribution, these observations on trait 1 will also allow the farmer to update her beliefs regarding μ2 . Since both traits 1 and 2 are present in the stacked variety, observations on trait 1 will allow the farmer to update her beliefs about the mean yields of the stacked variety.
9
The learning process that uses observations on trait one to update the beliefs about this same trait can be conceptualized as a standard Bayesian problem in which the prior is a normal univariate distribution. Consequently, the posterior distribution of μ1 will be distributed normally. Denoting this posterior as p1 ( μ1 ) , we have that: p1 ( μ1 ) ∼ N (θ11 ,V11 )
(7)
where the expected value θ11 and variance V11 can be calculated by the standard formula that corresponds to a Bayesian updating of the beliefs when the prior is a univariate normal distribution (Anderson). That is:
θ10
θ11 =
V11 =
+
nδ1
V σ 12 1 n + 2 0 V1 σ 1 0 1
1 1 n + 2 0 V1 σ 1
(8)
(9)
In equation (8) the farmer’s expected value of the posterior distribution will be a function of the expected value of the prior distribution, θ1o , the variance of that distribution, V1o , the known variance of yields, σ12 , her observed mean yield, δ1 , and the number of trials the farmer has had with the technology, n. We can see that if the average yield observed through experience is higher than the expected value of the prior distribution, the expected value of the beliefs regarding μ1 will be updated positively. If the contrary is true, the expected value of the beliefs will be updated negatively. Note that, although not explicitly modeled, the mean of the prior distribution θ1o may be influenced by common knowledge
10
of the technology outside of actual experience. Equation (9) shows the classical result that as the number of observations, n, increase, the variance of the distribution of the beliefs decrease, i.e.,
∂V11 ≤ 0 and that V11 ≤ V1o . This variance can be regarded as the uncertainty ∂n
faced by the farmer, an uncertainty that decreases as she learns through experience. The posterior distribution of μ2 can be deduced from the joint prior distribution, equation (5), together with the posterior distribution of μ1 , equation (7), in the following manner:
p1 ( μ2 ) ∼ ∫ p o ( μ2 | μ1 ) *p1 ( μ1 )d ( μ1 ) = g (θ 21 , V21 )
(10)
Where g is a distribution function with a expected value given by: θ 21 and a variance given by V21 . Taking into account the properties of the conditional distribution associated with a multivariate normal, we have that (see Appendix 1): C120 1 θ = θ + 0 (θ1 − θ10 ) V1 1 2
0 2
(11) 2
(C 0 ) 2 ⎛ C 0 ⎞ V = V − 120 + ⎜ 120 ⎟ *V11 V1 ⎝ V1 ⎠ 1 2
0 2
.
(12)
Definition 1: A technology is “ascendant” when θ11 ≥ θ10 and
∂θ11
> 0 , i.e., it is a ∂n technology for which new information has a non-negative effect on the mean of the distribution of the beliefs about the mean returns to the technology.
Lemma 1: When technology 1 is ascendant, and C12>0, then θ12 ≥ θ 2o and Lemma 2: When C12 > 0, then V21 ≤ V2o and
∂θ 21
∂n
>0 .
∂V21 0, then C12 ≤ C12o and
∂C121 and > the single trait variety, i.e, ∂n ∂n ∂n ∂n
Proposition 2 states that the uncertainty of the stacked variety will decrease more than the uncertainty of the single trait varieties. This proposition follows from equation (25), Lemma 2 and Lemma 3.
Proposition 3: For an ascendant technology, the incentive to adopt a stacked variety is increasing in the number of trials with a single trait variety.
15
Proof: The condition for the farmer to adopt a stacked variety is: sc
= EU s − EU c ≥ 0
and
s1
= EU s − EU1 ≥ 0
and
s2
= EU s − EU 2 ≥ 0
(26)
Looking at the impact of experience with trait 1 over the inequalities in (26), we have: ∂V (π ) ∂ sc ∂π s = − b ag s ∂n ∂n ∂n
(27)
∂ s1 ∂θ 1 ( p m ) 2 ∂V21 2( p m ) 2 ∂C121 = p m * ( 2 ) − b( + ) ∂n ∂n ∂n ∂n
(28)
∂ s2 ( p m ) 2 ∂V11 2( p m ) 2 ∂C121 ∂θ 1 ) = p m * ( 1 ) − b( + ∂n ∂n ∂n ∂n
(29)
B1
The first term on the RHS of equations (27)-(29) reflects the impact of experience on the difference between the expected profits of the stacked variety and the expected profits of the alternative seed. The second term reflects the impact of experience on the difference between the aggregate variance of the stacked variety and the aggregate variance of the alternative seed. Taking into account Proposition 2, the second term will always be positive. That is, the aggregate variance of the stacked variety will always decrease more than the aggregate variance of the alternative seed. On the other hand, the sign of the first term will depend on whether the technology is ascendant or not. With an ascendant technology, the beliefs regarding the mean yields of trait 1 are updated upwardly or do not change. As stated in Lemma 1 this implies that the mean of the beliefs regarding trait 2 are also updated upwardly or do not change. Thus, the first term of the RHS of equations (27)-(29) will be greater than or equal to zero.
16
Consequently, for an ascendant technology the impact of experience on s2
sc
,
s1
and
will be greater than zero, increasing the incentive to adopt a stacked variety.
Proposition 4: For a non-ascendant technology, the incentive to adopt a stacked variety increases with the number of trials with a single trait variety if the reduction of the difference between expected yields is lower than the reduction of the difference between the variance of the expected utility of stacked versus non stacked seeds.
For a non-ascendant technology, where the beliefs are updated downwards, we have that the first term of the RHS of equations (27)-(29) is lower than zero while the second term of these equations is higher than zero. Hence, the impact of experience on
sc
,
s1
and
s2
will be positive if the second term is higher in absolute value than the first term. This means that the incentive to adopt a stacked variety will increase with experience if the impact of experience on the difference between the variances of stacked versus non stacked seeds outweighs the negative impact of experience on the difference between the expected value of yields of stacked versus non stacked seeds. Propositions 3 and 4 demonstrate the potential for a sequential adoption pattern by showing that, under certain conditions, experience with single trait varieties will increase the chances that a farmer adopts stacked varieties. Note that if (26) holds for n=0; that is, if without a farmer observation on a single trait variety, the expected utility including both risk and uncertainty is higher for the stacked variety, then one would observe farmers jumping into the stacked variety directly rather than following a sequential adoption pattern through the single trait varieties.
Empirical Implementation
Econometrics of Sequential Adoption
17
Propositions 3 and 4 from the theoretical model show that, when the change in the expected value of yields is positive or if negative, higher than a certain threshold, the incentive to adopt a stacked variety will increase with the number of trials with the single trait variety Formally, they present the following testable probability statement: Pr( j* = s | ni ) > Pr( j* = s | n j )
for
ni > n j .6
(30)
A standard estimation method of testing (30) would imply the use of a logit or probit model, but the standard version of these models imposes a linear relationship between the independent variable, n, and the latent variable, EU j . Inspection of equations (8) and (9) shows that the latent variable, EU j depends non-linearly on n, the level of experience, with higher levels of n having lower effects on expected utility than do lower levels of n. Given this non-linearity in the role of experience, we estimate the adoption of a stacked variety using a semi-parametric specification.7 It is as follows:
yit = f (ait ) + ∑ αT ctT + vi + ε it
(31)
T
where, yit =1 if farm i adopts a stacked variety in year t, ait is experience, measured as the years that have passed since the first year of adoption of a GM crop, and ctT is a dummy variable for each chronological year, vi is a fixed effect, and εit is the standard
6
Testing equation (30) provides evidence consistent with propositions 3 and 4, while its rejection would imply that a technology is non-ascendant and proposition 4 does not hold. 7 An alternative specification, which can accommodate such non-linearity, would be to use survival analysis. Unfortunately, our data do not conform to the requirement of observing the characteristics of farmers at the beginning of the event, in this case the first introduction of GM seed traits in the mid 1990’s, which is necessary for a clean survival analysis estimate. 18
error term. The function f (ait ) has no predetermined parametric shape. The impact of experience occurs one year later: that is, if the farmer adopted GM corn for the first time in period t, ait would be equal to zero and ait +1 would be equal to one. In addition to an individual’s own experience, we expect that the expected utility of adopting a stacked variety may change over time. Including a time measure can capture rising levels of common information about the technology that a farmer access in addition to or as a substitute for their own experience. A key challenge in estimating the impact of experience arises with its potential endogeneity. Unobservable variables, such as the farmer’s ability, which could affect the adoption of stacked varieties, are also very likely to be correlated with early adoption. A possible solution implies the estimation of the likelihood of being an early adopter, which generates experience, as a function of the value of variables thought to cause early adoption such as land size and education, at the period at which the technology was first introduced. An exhaustive account of these variables, however, is very difficult to obtain using standard cross-sectional surveys, and is not feasible with our data. It is known (Heckman, Magnac) that the impact of a variable, such as experience, can also be consistently estimated through the inclusion of fixed effect parameters, which allow the researcher to isolate the role of experience from the role of other variables that vary by household but do not vary through time. However, a fixed effects regression can in turn introduce an “incidental parameters” problem. This problem arises in the context of maximum-likelihood estimation, and when the data include a small number of timeperiods as occurs in our data (Heckman). In order to estimate adoption as a function of experience while circumventing the incidental parameters problem, we use the Mundlak
19
device in which we run an estimation that controls for the average value of experience of each household. In the current setting, this amounts to controlling for the first year of adoption. Formally, since early adopters may possess characteristics that distinguish them from late adopters, it is possible for vi to be correlated with ait , which would introduce a bias in the estimation of f (ait ) . Following Mundlak, we can define vi as: vi = ∑ j =0 β j sij + zi 6
(32)
where sij is equal to one if the sum of the variable ait for farm i and over all the time periods is equal to j. While the fixed effects vi are usually expressed as a function of the average of the independent variable, we use a more flexible expression, in which there is one dummy for each value of this average experience level. This is equivalent to using a dummy for the first year of adoption, with two exceptions. The first is due to the exclusion of a dummy that is equal to one if the farmer adopted a GM crop in 2000. This dummy is excluded because using a set of dummies that sum one for each observation will imply a collinearity problem in the context of this semiparametric estimation. The second exception is related to the farmers who adopted for the first time in the last year of our data, 2006, or never adopted. These two groups of farmers are represented by the same dummy variable since they present the same value for the sum (and the average) of the variable ait , namely zero. Using the Mundlak formulation to account for potential correlation of vi and ait the adoption of stacked varieties can now be expressed as: yit = f (ait ) + ∑ αT ctT + ∑ j =0 β j sij + zi + ε it 6
(33)
T
20
where having controlled for early adoption through sij we can now assume that zi + ε it is not correlated with our variable of interest: ait . This semi-parametric specification offers the advantage of being highly flexible and of allowing us to know how the impact of experience changes as experience evolves. The estimation method follows Stock as shown in Appendix 2. Following Blundell and Duncan, we estimate confidence intervals for the non-parametric parameters using the bootstrap method. Additionally, the bootstrap replications allow us to correct for the bias, which characterizes nonparametric specifications.
Data: The empirical analysis uses data gathered from a random sample of 738 corn farmers from Minnesota and Wisconsin from lists that the National Agricultural Statistics Service (NASS) maintains for their work on the agricultural census and other USDA data collection activities. These farmers were surveyed in 2006, and the questionnaire included retrospective questions on their use of GM corn varieties covering every year back to 2000. The empirical estimation exploits the variation in time of their adoption of GM varieties in order to assess the existence of a sequential pattern of adoption. We use data that covers the period 2000-2006 that contain information regarding the type of variety adopted each year. 8
8
Note that prior to 2005 we cannot distinguish the use of stacked varieties from the use of more than one single trait variety. This distinction is relatively unproblematic though, because in years when we have data on both stacked and multiple variety use, multiple variety use was very rare. For example, in 2005, the use of more than one variety was 2%, while the use of stacked varieties was 23%. 21
Descriptive Statistics: Our first empirical illustration is to compare the timing of a farmer’s first year of GM adoption with the use of stacked varieties in 2005, shown in table 1. The likelihood of using stacked varieties in 2005 was twice as high (53% vs 27%) for farmers whose first year of GM adoption was in 2000-01 versus those adopting first in 2004-05. These descriptive results are consistent with a positive impact of experience through the reduction of uncertainty as well as with a technology that has proved to not be significantly worse than expected. Clearly, the results are not exclusive of other potential explanations. For one, this relationship might not imply causality because early adopters might have adopted the stacked varieties before accumulating any experience. Second, if there are variables correlated with early adoption, they might also shape farmer’s use of stacked varieties. The econometric results in the next section exploit the panel data to account for the time the stacked variety was adopted and to isolate the impact of previous experience with single trait seeds. Table 1 First year of Adoption and Type of Adoption in 2006 (Planned) First year of adoption
% of farms that use % of farms that use % of farms that do not use stacked varieties in 2006 single trait varieties in 2006 GM seeds in 2006
2000-01 2002-03 2004-05 2006
53.4 41.0 29.6 27.1
40.4 47.4 50.0 72.9
6.2 11.6 20.3 0.0
Total
41.4
48.2
10.4
Table 2 below shows the variables used in the semi-parametric estimation of equation (33) and the sample size of 4,157 observations. A farmer has adopted a stacked variety at 18% of the individual-year combinations in the data. On average across the dataset, 26% of
22
the farmers adopted a GM variety in 2000 while 40% of the farmers adopted GM for the first time in 2006 or never adopted. Semi-Parametric Estimation Results
Using the data summarized in table 2 we estimate the probability of adopting a stacked variety as expressed in equation (33). Since we do not impose any functional form for the impact of experience, it is possible to compare this impact for different groups.9 In order to assess how this impact changes with time we have estimated equation (33) for the whole sample as well as for a sub-sample that includes observations for the later period only. To assess how the impact of experience changes with human capital, we have estimated equation (33) for two sub-samples: the first one contains farmers with no college education, while the second one contains farmers with college education10.
9
If, instead, we had assumed, for example, a linear relationship, a low coefficient on experience might have meant that farmers learn fast and later experience does not matter but it could have also meant that experience does not matter at any point. This vagueness would make it hard to compare the impact of experience for different groups. Allowing the data to determine the shape of the relationships under analysis implies a more precise comparison between groups than if we had tried to model non-linearities in a parametric way such as with the use of quadratic terms. 10 When assessing how the impact changes with time we have not divided the sample in the early and late period because the range of the variable experience is too short in the early period. 23
Table 2 Descriptive Statistics Variable
Mean Std. Dev.
Adopted more than one trait (Yes=1, No=0)
0.18
0.38
Number of years since 1st year of adoption
1.22
1.81
Chronological year is 2002 (Yes=1, No=0)
0.17
0.38
Chronological year is 2003 (Yes=1, No=0)
0.17
0.38
Chronological year is 2004 (Yes=1, No=0)
0.17
0.38
Chronological year is 2005 (Yes=1, No=0)
0.16
0.37
Chronological year is 2006 (Yes=1, No=0)
0.16
0.36
Adopted a GM crop for 1st time in 2000 (Yes=1, No=0)
0.26
0.44
Adopted a GM crop for 1st time in 2001 (Yes=1, No=0)
0.04
0.19
Adopted a GM crop for 1st time in 2002 (Yes=1, No=0)
0.07
0.25
Adopted a GM crop for 1st time in 2003 (Yes=1, No=0)
0.08
0.27
Adopted a GM crop for 1st time in 2004 (Yes=1, No=0)
0.06
0.24
Adopted a GM crop for 1st time in 2005 (Yes=1, No=0)
0.09
0.28
Adopted a GM crop for 1st time in 2006 (Yes=1, No=0) or never adopted
0.40
0.49
yit ait
tt
sij
Number of observations
4157
Figure 1 shows the non-parametric component of the regression f(ait), estimated according to equation (33), for the whole sample (2000-2006 period) as well as for the 2004-2006 period. The bands indicate the 90% confidence intervals around the predicted value lines11. The whole period regression includes as intercept the dummy for the year
11
As is common in the non-parametric literature, we chose to show the significance intervals at a 90% confidence level because the flexibility that characterizes nonparametric regressions comes at the cost of estimators with higher variances. 24
2002, and the “later period” regression includes as intercept the dummy for the year 2005. The detailed results of both estimations are in Appendix 3. Figure 1 shows that, for the whole sample as well as for the later period, the likelihood of adopting more than one trait increases with experience. Experience is statistically significant for the whole period, since in this estimation there are some values of experience for which the confidence intervals do not intersect. These results are consistent with a positive impact of experience in reducing uncertainty and with Proposition 3 if a technology is ascendant or with the conditions stated in Proposition 4. On the other hand, the estimation for the later period presents a flatter slope and there are no values of experience for which the confidence intervals do not intersect, implying that experience is not a statistically significant predictor for the later period. Table 3 shows the change in the likelihood of adopting more than one trait as experience increases by one year, for the whole sample as well as for the 2004-2006 period. For the whole sample, the impact of experience is lower at the outset, begins to increase at 2 years of experience and slows down at 5 years of experience. The impact of one additional year of experience on the likelihood of adoption ranges between 1.09% and 3.4%. This magnitude can be considered economically significant given that the likelihood of adopting more than one trait for the individual-year combinations in the data is 18%, as shown in Table 2. As expected from Figure 1, Table 3 shows that the impact of experience is lower for the later period. Since the confidence interval of this difference lies in the positive range when the number of years of experience is equal to zero, three, four and five, we can assert that the difference in impacts between periods is statistically significant at the 95% level.
25
The flatter slope for own experience as well as the higher intercept in the later period likely reflect the higher degree of common knowledge about the technology, an issue we return to below. This first set of results supports the proposition that own experience shapes sequential adoption but diminishes in importance as own experience becomes less valuable as a way of acquiring information on the benefits of the technology.
Figure 1 Likelihood of adopting more than one GM trait for the whole sample and for the 2004-06 period. 90% confidence intervals
70 60 50 40 30 20 10 0
1
2
3
4
5
6
Number of years since 1st year of adoption
Whole Sample
Later Period
Table 3 Change in the likelihood of adopting more than one trait as experience increases by one year. Whole Sample (2000-2006) vs Later Period (20042006) Experience *
Whole sample Later Period A B 0 1.09 0.75 1 1.80 1.32 2 2.71 1.88 3 3.37 1.97 4 3.27 1.55 5 2.53 1.06 Total 14.77 8.54 * Years since the first year of adoption **Estimated using bootstrap
Difference 95 % Confidence Interval A-B of the Difference** 0.34 0.02 0.63 0.48 -0.07 0.99 0.82 -0.03 1.55 1.40 0.40 2.18 1.73 0.78 2.46 1.47 0.64 2.09 6.24 2.29 9.55
26
Next in Figure 2, we consider the non-parametric component of the estimation for two sub-samples: one that contains farms whose operator has no college education and the other farms whose operator has a college degree or a higher educational level. In this figure, both functions include as intercept the dummy for the year 2002. For the sake of visual clarity, the confidence intervals are not included in this figure, but are available in Appendix 3. For both sub-samples, the 90% confidence interval that corresponds to six years of experience does not intersect with the confidence interval that corresponds to zero years of experience, implying that experience is statistically significant for both the college-educated and non college-educated farmers at the 90% level of confidence. Table 4 shows the change in the likelihood of adopting more than one trait as experience increases by one year. The results show that the impact of experience for college-educated farmers is higher than the impact for non college educated farmers at low levels of experience and lower at higher levels of experience. The difference between the impact of experience for these two sub-samples is negative and statistically different than zero when farmers have zero experience while being positive and statistically different from zero for farmers with four and five years of experience. Overall these results indicate that college- educated farmers tend to learn faster from own experience than less educated farmers.
27
Figure 2 Likelihood of adopting more than one GM trait for the non college and for the college educated farmers. 90% confidence intervals
45 40 35 30 25 20 15 10 5 0
1
2
3
4
5
6
Number of years since 1st year of adoption
College Educated
Non College Educated
Table 4 Change in the likelihood of adopting more than one trait as experience increases by one year. Non College Educated vs College Educated Farmers. Experience *
Non College Educated A 0 0.70 1 1.24 2 2.16 3 3.31 4 4.06 5 3.89 Total 15.35 * Years since the first year of adoption **Estimated using bootstrap
College Educated B 1.46 2.14 2.74 2.80 2.20 1.38 12.71
Difference 95 % Confidence Interval A-B of the Difference** -0.76 -1.20 -0.32 -0.90 -1.55 -0.18 -0.58 -1.66 0.59 0.51 -1.01 2.13 1.86 -0.04 3.66 2.51 0.35 4.09 2.65 -3.88 9.58
In Figure 3, we use the time variables to investigate how aggregate levels of information affect the probabilities of individuals adopting the technology. It depicts the coefficients associated with the year dummies of the semi-parametric regression for the whole sample, i.e., α t in specification (33). The coefficients on the year dummies have 28
increased through time, with the diffusion of the technology, implying increases in the base levels of stacked variety adoption one would predict from a typical S-curve adoption pattern. Comparing the coefficient values in Figure 3 to Figure 1 provides insight into the impact of direct experience with a technology (Figure 1) with indirect experience (Figure 3). The results from Figure 1, of a positive impact of experience for the whole sample provide evidence of a sequential pattern adoption. However, the results shown in Figure 3 of a higher likelihood to adopt as time evolves, combined with the results on Table 3, which show that the impact of experience is lower for the later period, supports the notion that as time evolves and the uncertainty with a new technology dissipates, farms are more likely to jump directly to the use of stacked varieties and own experience loses some of its explanatory power.
Figure 3 Likelihood of adopting more than one GM trait as a function of the chronological year. 90% confidence interval 30.0 25.0 20.0 15.0 10.0 5.0 0.0 2002
2003
2004
2005
2006
Year
29
In sum, the results from the semi-parametric estimation of the adoption of stacked crop varieties shows a sequential adoption process in which the initial experience with a single trait variety reduces the uncertainty and increases the probability of adopting a stacked variety. The estimates also show that farmers with high levels of education tend to learn faster from own experience and require fewer years of experience in order to adopt a stacked variety. In addition, the estimates show that with time, as the general level of knowledge in society about stacked varieties increases, this knowledge reduces uncertainty with the technology. Accordingly, over time the benefits of own experience and incentives for sequential adoption decline.
Conclusions
GM corn seed companies have innovated continuously with the introduction of new traits and, more recently, with the creation of stacked varieties, which combine more than one trait. In spite of its potential importance, no previous studies had examined the determinants of adoption of stacked varieties, particularly the learning dynamics and potential for sequential adoption in which farmers moved from use of GM varieties with a single trait to adoption of varieties with stacked traits. Our article fills this gap in the literature. The Bayesian modeling framework is to our knowledge the first that models a sequential adoption process for a packaged good with correlated component technologies. It demonstrates the conditions under which experience with a single component of a package technology such as traits in corn seeds plays a role in reducing uncertainty regarding the profitability of the packaged (stacked) technology. As a consequence of this
30
learning process, farmers’ uncertainty associated with adopting a stacked variety will be lower conditional on having previously adopted one of the traits included in the seed. The empirical results indicate that early adopters of GM will be more likely to adopt stacked varieties and benefit from the higher yields they offer. The empirical evidence presented above shows that the likelihood of adopting a stacked variety increases with experience, measured as the number of years that have passed since the first year of adoption of a GM technology. The results also show that the impact of experience decreases with time as the technology diffuses, and that farmers are more likely to jump directly to the use of stacked varieties in the later years. They also show that more educated farmers tend to learn faster from own experience. The sequencing of adoption of a package technology has important implications for both the introduction of new technologies to the market and the industrial organization of the market for new technologies. In terms of the introduction of new technologies, these results suggest strong incentives for technology sellers or promoters to help early adoption of component parts of a stacked technology as a strategy for eventual adoption of the package technology. In contrast, efforts to push adoption of stacked technologies that ignore the uncertainty that drives sequential adoption processes may run into difficulty. The sequential nature of adoption may drive differential pricing strategies among technology sellers and this could have profound implications for competition and the ability of small producers to compete. These industrial organization questions are left for future research.
31
References
Alexander, C., J. Fernandez Cornejo and R. Goodhue. 2001. “Determinants of GMO Use: A Survey of Iowa Maize-Soybean Farmers’ Acreage Allocation.” In: V. Santinello, R.E. Evenson, and D. Zilberman eds. Market Development for Genetically Modified Foods, Cabi Publishing. Alexander, C and T . Van Mellor. 2005. “Determinants of Corn Rootworm Resistant Corn Adoption in Indiana”. Agbioforum 8: 197-204. Alexander, C. 2006. “Farmer Decisions to Adopt Genetically Modified Crops.” CAB Reviews: Perspectives in Agriculture, Veterinary Science, Nutrition and Natural Resources 1: 1-9. Anderson, J. 1977. Agricultural Decision Analysis, Iowa City: Iowa University Press. Barham, B., J. Foltz, D. Jackson Smith and S. Moon. 2004. “The Dynamics of Agricultural Biotechnology Adoption: Lessons from rbst use in Wisconsin, 19942000.” American Journal of Agricultural Economics, 86: 61-72. Barham, B., M.R. Carter, and W. Sigelko. 1995. “Agro-Export Production and Peasant Land Access: Examining the Dynamics between Adoption and Accumulation.” Journal of Development Economics 46:85-107. Below, F.E., M. Uribelarrea, M. Ruffo, S.P. Moose and A.W. Becker. 2007. “Triplestacks, Genetics and Biotechnology in Improving Nitrogen Use of Corn.” Proc. North Central Extension-Industry Soil Fertility Conference, 23:5-13. Blundell, R. and A. Duncan. 1998. “Kernel Regression in Empirical Microeconomics.” The Journal of Human Resources, 33: 62-87. Byerlee, D. and E. Hesse de Polanco. 1986. “Farmers' Stepwise Adoption of Technological Packages: Evidence from the Mexican Altiplano.” American Journal of Agricultural Economics, 68: 519-527. Carletto, C., A. De Janvry, and E. Sadoulet. 1999. “Sustainability in the Diffusion of Innovations: Smallholder Nontraditional Agro-Exports in Guatemala.” Economic Development and Cultural Change, 47: 345-370. Chen, L. and Buttel, F.H. 2000. “Dynamics of GMO Adoption among Wisconsin Farmers”. PATS Staff Paper Series. Paper # 4. Conley, T.G. and C.R. Udry. 2004. “Learning About a New Technology: Pineapple in Ghana.” Economic Growth Center, Yale University. Working Paper # 817.
32
Dong, D. and A. Saha. 1998. “He Came He Saw and He Waited: an Empirical Analysis of Inertia in Technology Adoption.” Applied Economics 30: 893-90. Feder, G., R. Just, and D. Zilberman. 1985. “Adoption of Agricultural Innovations in Developing Countries: A Survey.” Economic Development and Cultural Change 33 : 255-298. Feder, G. and R. Slade. 1984. “The Acquisition of Information and the Adoption of Technology.” American Journal of Agricultural Economics, 66: 312-320. Feder, G. and G. OMara. 1981. “Farm Size and the Diffusion of Green Revolution Technology.” Economic Development and Cultural Change, 30: 59-76. Fernandez Cornejo, J., A. Corrine and R. Goodhue. 2002. “Dynamic Diffusion with Disadoption: The Case of Crop Biotechnology in the USA.” Agricultural and Resource Economics Review, 31: 112-126. Fernandez Cornejo, J., S. Daberkow and W. McBride. 2001. “Decomposing the Size Effect on the Adoption of Innovations: Agrobiotechnology and Precision Agriculture.” Agbioforum, 4: 124-136. Foltz, J.D. and H. H. Chang. 2002. “The Adoption and Profitability of rBST on Connecticut Dairy Farms.” American Journal of Agricultural Economics, 84: 1021-1032. Foster, A. and M. Rosenzweig. 1995. “Learning by Doing and Learning From Others: Human Capital and Technical Change in Agriculture.” Journal of Political Economy, 103:1176-1209. Griliches, Z. 1957. "Hybrid Corn: An Exploration in the Economics of Technological Change." Econometrica 25:501-22. Hategekimana, B and M. Trant. 2002. “Adoption and Diffusion of New Technology in Agriculture: Genetically Modified Corn and Soybeans.” Canadian Journal of Agricultural Economics, 50: 357-371. Heckman, J.J. 1981. “The Incidental Parameters Problem and the Problem of Initial Conditions in Estimating a Discrete Time-Discrete Data Stochastic Process.” In: C.F. Manski and D. McFadden eds. 1981. Structural Analysis of Discrete Data with Econometric Applications, Cambridge, MA: MIT Press. Horowitz , J and E. Lichtenberg. 1993. “Insurance, Moral Hazard, and Chemical Use in Agriculture.” American Journal of Agricultural Economics, 75: 926-935.
33
Hubbell, B.J. M.C. Marra and G.A. Carlson. 2000. “Estimating the Demand for a New Technology: Bt Cotton and Insecticide Policies.” American Journal of Agricultural Economics, 82:118–32. Huffman, W. 1974. “Decision Making: the Role of Education.” American Journal of Agricultural Economics, 56: 85-97. Kahnna, M. 2001. “ Sequential Adoption of Site-Specific Technologies and its Implications for Nitrogen Productivity: A Double Selectivity Model.” American Journal of Agricultural Economics, 83: 35-51. Leathers, H and M. Smale. 1991. “A Bayesian Approach to Explaining Sequential Adoption of Components of a Technological Package.” American Journal of Agricultural Economics, 73: 734-742. Lindner, R. 1980 “Farm Size and the Time Lag to Adoption of a Scale Neutral Innovation.” Adelaide: University of Adelaide. Lindner, R., A. J. Fischer and P. Pardey. 1979. “The Time to Adoption.” Economic Letters 2: 187-90. Lindner, R., Pardey, P. and F. Jarrett. 1982. “Distance to Information Source and the Time Lag to Early Adoption of Trace Element Fertilizer.” Australian Journal of Agricultural Economics, 26: 100-113. Magnac, T. 2000. “Subsidized Training and Youth Employment: Distinguishing Unobserved Heterogeneity from State Dependence in Labour Market Histories.” The Economic Journal, 110: 805-837. Meyer, J., 1987. “Two-Moment Decision Models and Expected Utility Maximization.” The American Economic Review, 77: 421-430. Mitchell, P., T.M. Hurley and M.E. Rice. “How Bt Corn Affects the Distribution of Harvested Yield.” Mundlak, Y. 1978. “On the Pooling of Time Series and Cross Section Data.” Econometrica 46: 69-85. Rogers, E. 1986. Diffusion of Innovation, New York: The Free Press. Stock, J. 1989. “Nonparametric Policy Analysis.” Journal of the American Statistical Association, 84: 62-87. Stoneman, P. 1981. “Intra-Firm Diffusion, Bayesian Learning and Profitability” The Economic Journal, 91: 375-388.
34
Tsur,Y., S. M. Sternberg and E. Hochman. 1990. “Dynamic Modeling of Innovation Process Adoption with Risk Aversion and Learning.” Oxford Economic Papers, 42:336-355. Useche, P., B. Barham, and J. Foltz. 2009. “Integrating Technology Traits and Producer Heterogeneity. A Mixed Multinomial Model of GM Adoption.” American Journal of Agricultural Economics, 91: 444-461.
35
Appendix 1: Derivation of equations (11) and (12)
As expressed in equation (10) in the text, the posterior distribution can be derived from: p1 ( μ 2 ) = ∫ p o ( μ 2 | μ1 ) *p1 ( μ1 )d ( μ1 ) = g (θ 21 , V21 )
(1’)
which means that the posterior distribution of μ2 is based on the prior conditional distribution of μ2 given μ1 , as well as the posterior distribution of μ1 . We define the mean of p o ( μ2 | μ1 ) as θ 20/1 and the variance as V20/1 . By the characteristics of a bivariate normal, and taking (5) and (6) into account, we have that:
θ 20/1 = θ 20 +
C120 ( μ1 − θ10 ) V10
(2’)
V20/1 = V20 −
(C120 ) 2 V10
(3’)
Using the law of iterated expectations on (2’) and taking (7) into account, we have that: C0 (4’) θ 21 = θ 20 + 120 (θ11 − θ10 ) V1 which is equation (11) in the text. Using the law of iterated expectations, we can express the variance of the posterior of μ2 as: V21 = E (V20/1 ) + V (θ 20/1 ) μ1
(5’)
μ1
where the first term is equal to the expected value of V20/1 over the posterior distribution of
μ1 , and the second term is the variance of θ 20/1 over the posterior distribution of μ1 . This implies that:
2
(C 0 ) 2 ⎛ C 0 ⎞ V = V − 120 + ⎜ 120 ⎟ V11 V1 ⎝ V1 ⎠ 1 2
0 2
(6’)
where the first term is equal to V20/1 in (3’) and the second term is the variance of θ 20/1 . Equation (6’) is equation (12) in the text.
1.2 Derivation of equation (13)
36
The covariance between μ1 and μ2 satisfies the usual formula: C12i = E i ( μ1μ 2 ) − θ1iθ 2i
(7’)
Using the law of iterated expectations, we have that: E1 ( μ1μ 2 ) = ∫ μ1θ 20/1 p1 ( μ1 ) d ( μ1 )
(8’)
Replacing (2’) in (8’), we have:
E1 ( μ1μ 2 ) = θ 20θ11 +
C120 1 C120 0 1 E ( μ μ ) − θ1 θ1 1 1 V10 V10
(9’)
Since E1 ( μ1 μ1 ) = V11 + θ11θ11 , and taking (7’) as well as (9’) into account, C121 can be expressed as:
C121 = θ 20θ11 +
C120 1 1 1 C120 0 1 1 1 (V1 + θ1θ1 ) − 0 θ1 θ1 − θ 2θ1 V10 V1
(10’)
Using (4’) to replace θ 21 in (10’), we have that: C121 =
C120 V11 V10
(11’)
Equation (11’) is equation (13) in the text.
37
Appendix 2: Semi-parametric Estimation Method This appendix describes the method used in the estimation of the semi-parametric regressions. As specified in (33) we run the following regression: yit = f (ait ) + ∑ αT ctT + ∑ j =0 β j sij + zi + ε it 6
T
Setting dit = [ ct si ] , A = [α β ] and rit = zi + ε it Where ct = [c2002t c2003t .....c2006t ] , si = [ si 0 si1....si 6 ] , βi = [ β 0 β1....β 6 ]
α t = [α 2002 α 2003 .....α 2006 ] and
We have yit = f (ait ) + A ' dit + rit Taking expected values with respect to e and creating a new index j, which goes from 1 (when i=1 and t=1) to nt, we have that: y j − E ( y j | a j ) = A '(d j − E (d j |a j )) + rit − E (rit | a j )
From this expression it is easy to see that A can be estimated as (Stock(1989)): n
n
j =1
j =1
A = (∑ (d j − E (d |a j ))( d j − E ( d |a j )) ') −1 (∑ (d j − E (d |a j ))( y j − E ( y j |a j )))
Where
∑ k(
ak − a j
∑ k(
ak − a j
) yk h E ( y |a j ) = a −a ∑k k ( k h j ) k
)d k h E (d |a j ) = a −a ∑k k ( k h j ) k
Consequently, the non parametric part can be estimated as:
f (a j ) = E ( y |a j ) − A ' E (d |a j )
38
Appendix 3 Semi-parametric Estimation Results
Table A1 shows the results of the estimation of equation (33), for the whole sample and for the later period. The expected value of the non parametric part and the coefficients of the linear part of the regression have been multiplied by 100. Table A1
Semiparametric Regression Estimates for the Whole Sample and for the Later Period Dependent Variable: 1 if the farmer uses more than one GM trait
Whole Sample Years since the first year of adoption 0 1 2 3 4 5 6
Later Period: 2004-2006
Expected 90% Confidence Interval value 23.0 24.1 25.9 28.6 32.0 35.2 37.8
18.8 19.8 21.5 24.4 27.6 30.7 33.0
27.5 28.6 30.4 33.0 36.4 39.7 42.4
Expected value
90% Confidence Interval
38.99 39.75 41.07 42.95 44.92 46.47 47.53
32.94 33.82 35.43 37.55 39.63 41.30 42.36
44.90 45.59 46.63 48.51 50.27 51.79 52.71
Coefficient 90% Confidence Interval Coefficient 90% Confidence Interval Control Variables First year of adoption * 2001 2002 2003 2004 2005 2006
-8.4 -10.4 -9.8 -18.0 -21.6 -31.0
-18.6 -18.6 -17.4 -23.8 -26.6 -35.5
3.3 -2.6 -2.9 -12.2 -16.8 -26.5
-7.48 -8.39 -7.61 -14.91 -24.93 -44.30
-21.64 -18.59 -19.23 -24.77 -32.96 -50.00
7.40 2.35 3.68 -5.29 -17.21 -38.48
1.8 6.8 9.0 16.2 22.1
0.6 5.0 7.0 14.0 19.3
3.0 8.6 11.0 18.6 24.8
8.04 14.76
5.26 11.67
10.83 17.49
Chronological year ** 2002 2003 2004 2005 2006
* The dummy that equals one if the first year of adoption was on 2000, or before 2000, is not inculded in the estimation. Thus, the coefficients represent the impact relative to those who adopted for the first time in 2000 or before ** The dummy that equals one if the chronological year is 2001 is not included in the estimation
39
Table A2 shows the results of the estimation of equation (33), for two subsamples: farmers with a college degree or a higher educational level and farmers with no college degree. The expected value of the non parametric part and the coefficients of the linear part of the regression have been multiplied by 100. Table A2
Semiparametric Regression Estimates for the College Educated and for the Non College Educated Dependent Variable: 1 if the farmer uses more than one GM trait College Educated Non College Educated Years since the first year Expected Confidence Interval Expected Confidence Interval value value of adoption 0 1 2 3 4 5 6
Control Variables First year of adoption * 2001 2002 2003 2004 2005 2006
25.9 27.3 29.5 32.2 35.0 37.2 38.6
19.6 21.2 23.4 26.2 29.0 31.1 32.4
32.0 33.5 35.5 38.3 40.9 43.5 44.8
16.892 17.589 18.832 20.990 24.301 28.361 32.247
10.583 11.307 12.564 14.996 18.193 21.798 25.166
23.076 23.723 25.115 26.758 30.304 34.572 39.056
Coefficient Confidence Interval Coefficient Confidence Interval -14.1 -10.3 -10.7 -24.2 -24.8 -36.1
-29.1 -20.9 -21.7 -32.5 -31.7 -42.2
0.3 0.1 -0.3 -16.4 -18.0 -30.1
2.5 -12.6 -5.8 -8.4 -15.5 -23.4
-16.8 -24.7 -15.7 -16.0 -21.9 -29.3
19.5 -1.1 4.7 -0.6 -8.2 -17.5
2.7 9.2 12.3 20.6 25.3
0.9 6.4 8.9 16.2 21.1
4.4 12.0 15.7 25.1 29.4
1.3 4.9 7.0 13.1 19.8
0.0 2.4 4.5 9.8 16.0
2.5 7.2 9.5 16.2 23.4
Chronological year ** 2002 2003 2004 2005 2006
* The dummy that equals one if the first year of adoption was on 2000, or before 2000, is not included in the estimation. Thus, the coefficients represent the impact relative to those who adopted for the first time in 2000 or before ** The dummy that equals one if the chronological year is 2001 is not included in the estimation
40